我想使用 Tensorflow 对象检测 API 创建一个模型来检测信用卡中的卡号。因此，我准备了大约 50000 张卡片用于训练和 15000 张卡片用于验证的卡片数据集。我的模型是 SSD_Mobilenet_V1_0.25_224，我运行 280K 步骤的训练。一切看起来都很好，我的 total_training_loss 低于 1，约为 0.8，我的 validation_classification_loss 为 0.7，validation_localication_loss 约为 0.02，average_persion > 为 1.0。这是我的情节，它们似乎很好:
这是我的配置:
# SSD with Mobilenet v1 configuration for MSCOCO Dataset.
# Users should configure the fine_tune_checkpoint field in the train config as
# well as the label_map_path and input_path fields in the train_input_reader and
# eval_input_reader. Search for "PATH_TO_BE_CONFIGURED" to find the fields that
# should be configured.
model {
ssd {
num_classes: 1
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
x_scale: 10.0
height_scale: 5.0
width_scale: 5.0
}
}
matcher {
argmax_matcher {
matched_threshold: 0.5
unmatched_threshold: 0.5
ignore_thresholds: false
negatives_lower_than_unmatched: true
force_match_for_each_row: true
}
}
similarity_calculator {
iou_similarity {
}
}
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.1
max_scale: 0.95
aspect_ratios: 1.0
aspect_ratios: 3.0
aspect_ratios: 6.0
aspect_ratios: 9.0
aspect_ratios: 10.32
aspect_ratios: 11.96
aspect_ratios: 12.06
aspect_ratios: 13.9
aspect_ratios: 12.96
aspect_ratios: 14.71
aspect_ratios: 13.65
aspect_ratios: 16.27
aspect_ratios: 17.73
aspect_ratios: 18.68
aspect_ratios: 16.74
aspect_ratios: 14.91
aspect_ratios: 13.33
aspect_ratios: 10.67
aspect_ratios: 10.5
aspect_ratios: 10.26
aspect_ratios: 10.81
aspect_ratios: 10.31
aspect_ratios: 11.05
aspect_ratios: 11.52
aspect_ratios: 11.0
aspect_ratios: 12.58
aspect_ratios: 12.12
aspect_ratios: 12.8
aspect_ratios: 13.97
aspect_ratios: 13.34
aspect_ratios: 13.45
}
}
image_resizer {
fixed_shape_resizer {
height: 300
width: 500
}
}
box_predictor {
convolutional_box_predictor {
min_depth: 0
max_depth: 0
num_layers_before_predictor: 0
use_dropout: false
dropout_keep_probability: 0.8
kernel_size: 1
box_code_size: 4
apply_sigmoid_to_scores: false
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
}
feature_extractor {
type: 'ssd_mobilenet_v1'
min_depth: 16
depth_multiplier: 0.25
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
loss {
classification_loss {
weighted_sigmoid {
}
}
localization_loss {
weighted_smooth_l1 {
}
}
hard_example_miner {
num_hard_examples: 3000
iou_threshold: 0.99
loss_type: CLASSIFICATION
max_negatives_per_positive: 3
min_negatives_per_image: 0
}
classification_weight: 1.0
localization_weight: 1.0
}
normalize_loss_by_num_matches: true
post_processing {
batch_non_max_suppression {
score_threshold: 1e-8
iou_threshold: 0.6
max_detections_per_class: 100
max_total_detections: 100
}
score_converter: SIGMOID
}
}
}
train_config: {
batch_size: 64
optimizer {
rms_prop_optimizer: {
learning_rate: {
exponential_decay_learning_rate {
initial_learning_rate: 0.004
decay_steps: 5000
decay_factor: 0.95
}
}
momentum_optimizer_value: 0.9
decay: 0.9
epsilon: 1.0
}
}
fine_tune_checkpoint: "/home/shayantabatabaei/Projects/CardScanner/NeuralNetwork/trainer/model/mobilenet_v1_0.25_224.ckpt"
from_detection_checkpoint: false
# Note: The below line limits the training process to 200K steps, which we
# empirically found to be sufficient enough to train the pets dataset. This
# effectively bypasses the learning rate schedule (the learning rate will
# never decay). Remove the below line to train indefinitely.
num_steps: 450000
data_augmentation_options {
random_horizontal_flip {
}
}
data_augmentation_options {
ssd_random_crop {
}
}
}
train_input_reader: {
tf_record_input_reader {
input_path: "/home/shayantabatabaei/Projects/CardScanner/NeuralNetwork/dataset/images/train.record"
}
label_map_path: "/home/shayantabatabaei/Projects/CardScanner/NeuralNetwork/trainer/labelmap.pbtxt"
}
eval_config: {
num_examples: 14000
# Note: The below line limits the evaluation process to 10 evaluations.
# Remove the below line to evaluate indefinitely.
# max_evals: 10
num_visualizations: 50
}
eval_input_reader: {
tf_record_input_reader {
input_path: "/home/shayantabatabaei/Projects/CardScanner/NeuralNetwork/dataset/images/test.record"
}
label_map_path: "/home/shayantabatabaei/Projects/CardScanner/NeuralNetwork/trainer/labelmap.pbtxt"
shuffle: false
num_readers: 1
}
一切似乎都很好，但是当我将模型导出为 tflite 格式并在移动设备上使用时，它没有找到任何卡号。 这是我的数据集的示例:
我的模型似乎过度拟合了吗？我该如何解决这个问题？
谢谢!
最佳答案
最后我找到了解决方案，我将我的配置文件更改为此并添加更多的aspect_ratios，它会增加我的模型在框预测层中的权重，并删除冗余的aspect_ratios。
这是我的配置文件:
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.1
max_scale: 0.95
aspect_ratios: 1.0
aspect_ratios: 1.5
aspect_ratios: 2.0
aspect_ratios: 2.5
aspect_ratios: 3.0
aspect_ratios: 3.5
aspect_ratios: 4.0
aspect_ratios: 4.5
aspect_ratios: 5.0
aspect_ratios: 5.5
aspect_ratios: 6.0
aspect_ratios: 6.5
aspect_ratios: 7.0
aspect_ratios: 7.5
aspect_ratios: 8.0
aspect_ratios: 8.5
aspect_ratios: 9.0
aspect_ratios: 9.5
aspect_ratios: 10.0
aspect_ratios: 10.5
aspect_ratios: 11.0
aspect_ratios: 11.5
aspect_ratios: 12.0
aspect_ratios: 12.5
aspect_ratios: 13.0
aspect_ratios: 13.5
aspect_ratios: 14.0
aspect_ratios: 14.5
aspect_ratios: 15.0
aspect_ratios: 15.5
aspect_ratios: 16.0
aspect_ratios: 16.5
aspect_ratios: 17.0
aspect_ratios: 17.5
aspect_ratios: 18.0
aspect_ratios: 18.5
aspect_ratios: 19.0
aspect_ratios: 19.5
aspect_ratios: 20.0
aspect_ratios: 20.5
aspect_ratios: 21.0
}
}
我遇到的另一个问题是我没有规范化 Android 代码中的输入，因此根据 this file SSD_MOBILENET 将标准化范围 [-1,1] 之间的输入，因此我像这样更改我的 android 代码:
@Override
protected void addPixelValue(int pixelValue) {
imgData.putFloat(normalizeValue((pixelValue >> 16) & 0xFF));
imgData.putFloat(normalizeValue((pixelValue >> 8) & 0xFF));
imgData.putFloat(normalizeValue(pixelValue & 0xFF));
}
private float normalizeValue(float value) {
return value * (2 / 255.0f) - 1.0f;
}
终于成功了!
关于即使损失很低，Tensorflow 对象检测 API 也无法工作，我们在Stack Overflow上找到一个类似的问题： https://stackoverflow.com/questions/59138588/